[Core] Saturate float-to-integer cast against NaN/Inf/out-of-range in reference::detail::convert

[axinging]

Details:

ov::reference::detail::convert<TI, TO>() cast floats to integral types with a raw static_cast, which is undefined behaviour per [conv.fpint] for NaN, ±Inf, and any finite value that doesn't fit the destination type (e.g. static_cast<int8_t>(300.0f)). An attacker-supplied model can trigger this via a Constant → Convert pair evaluated during constant folding (or at runtime).

Fixed by adding a saturating overload: NaN → 0, out-of-range (including ±Inf) → destination min/max. A second, independent instance of the same bug in the JIT-backed float/float16 → int8_t specializations (src/core/reference/src/op/convert.cpp) — which bypass detail::convert entirely via unguarded hand-written assembly — is fixed separately: since no clamped JIT kernel exists for these type pairs, they now call Converter<TI,TO>::apply<Clamp<TI,TO>>(...) directly, matching the existing convert<int32_t, float16> pattern.

Also closes two related gaps found while implementing the above: float16/bfloat16 sources weren't recognized as floating-point by the NaN guard (std::is_floating_point<float16> is false — they're class types with an implicit operator float()), and the saturating overload was incorrectly being instantiated for TO = bool (MSVC C4804, and a [conv.bool] violation — NaN/Inf must map to true, not false).

This is a behaviour change, not just a UB fix: e.g. 300.0f → int8_t now deterministically saturates to 127 instead of wrapping to whatever a given compiler/platform produced for the UB cast.

Tests:

• src/core/tests/eval.cpp: evaluate_convert_f32_nan_inf_to_i32_no_ub, evaluate_convert_f32_out_of_range_to_i8_saturates, evaluate_convert_f16_out_of_range_to_i8_saturates.
• src/core/tests/pass/constant_folding.cpp, constant_folding.const_convert: updated two sub-cases that hardcoded the old UB-dependent wraparound result to the new saturated values.

TODO:

Add jit_convert_vec<float, int8_t, true> / jit_convert_vec<float16, int8_t, true> (AVX2, clamp via vminps/vmaxps against [-128.0f, 127.0f] before vcvttps2dq, plus an explicit NaN mask+blend since vminps/vmaxps return the second operand on NaN, not 0) to restore JIT/AVX2 speed for these two conversions instead of the current scalar fallback. While at it, remove the now-dead clamp=false specializations at convert.cpp:168 and :187 if no longer needed. Requires new unit tests covering NaN/±Inf/boundary values before enabling.

Tickets:

• CVS-194588

AI Assistance:

• AI assistance used: yes
• Bug root cause was found by AI.